acceptodds
Desk rejected at ICLR 2027

SELECTIVE INVARIANCE UNDER STRATEGIC INFORMATION MANIPULATION

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

An agent that buys on a user's behalf reads a seller's description, contract terms and framing, then accepts the offer or walks away. Such a rule must not move when the seller merely re-presents a fixed transaction, yet must still move when the transaction genuinely changes. The latent state is unavailable at inference, so the rule cannot tell the two apart by inspection; it must be built so that one class of change cannot reach it. We call this selective invariance and ask when it is attainable. Given a specified admissible channel and observable representation, comparing the displacements that channel can produce with the direction that carries utility gives a criterion that sorts channels into three classes: invariance is free, invariance costs enumerating a completion set, or invariance is unattainable and the channel can only be estimated. Existing structural defenses show how to isolate particular flows; our contribution is an ex ante test of whether exact model-independent closure exists for a specified channel, not a detector that discovers arbitrary free-text manipulations. Because an LLM buyer is an opaque model rather than one we design, a guarantee valid for every decision function requires exact invariance, which no prompt-level defense provides. We measure both halves on paired renderings of the same hidden transaction, across three second-hand marketplace and bargaining benchmarks and seven model families. Presentation alone raises harmful acceptance by even for a buyer that accepts no bad offer under the honest rendering, and a safety prompt's protection varies seventeen-fold across families. Structural closure cuts harmful acceptance from to at essentially unchanged valid acceptance, holds the closed channels at exactly zero, and still reproduces a hidden-state oracle's decision under genuine concessions on all held-out negotiations. Applied unchanged to tool-using agents inside AgentDojo, with their suite, their injections and their verdicts, the criterion predicts in advance which defenses can be exact: ours takes attacker success from to exactly on the channel it closes, at no cost in task utility, while the published datamarking baseline reaches .

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
selective invariance, strategic manipulation, LLM agents, multi-agent system, agent safety

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nfYbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how purchasing agents can resist misleading seller messages while still responding to real improvements in a deal. It proposes a three-way framework for deciding which types of manipulation can be blocked using the available information. The method removes irrelevant information, evaluates uncertain contract terms using their worst possible outcomes, and uses conservative estimates for unverified item quality.

Critical strengths

The paper addresses a practical problem and provides a useful framework for organizing manipulation risks. Its paired tests keep the underlying transaction fixed while changing its presentation, helping isolate the effects of manipulation. Tests with genuine concessions also check whether the defense remains responsive to better offers. The experiments cover bargaining datasets and AgentDojo.

Critical weaknesses

The existing propositions separately address invariance, monotonicity under set inclusion, and non-identifiability. However, it is unclear how they jointly establish a complete three-way classification. The authors should clarify this connection, since the classification is the paper’s main claimed theoretical contribution.

The AgentDojo experiments lack a comparison with existing structural defenses such as CaMeL. Including such baselines under the same attack assumptions and information access would help assess the proposed method’s benefits for safety and task performance.

Questions for authors

See the Critical Weaknesses.

3 comments

We revised Section 4.1 and the proofs to clarify how the three propositions relate to the classification. We also added a direct CaMeL comparison.

1. How do the propositions support the three-way classification?

We added Corollary 1 and clarified its scope.

For a specified channel and observable representation, let Δk(m)\Delta_k(m) be the admissible displacement set and let gg be the local utility gradient. At each observation, exactly one of three conditions holds:

(a) Every displacement has zero inner product with gg.

(b) All inner products are nonpositive, with at least one negative.

(c) At least one inner product is positive.

These conditions form an exhaustive partition of local directional signs. They do not, by themselves, prove that every channel admits one of three defense implementations.

Each operational guarantee requires further assumptions. Proposition 1 concerns exact representation invariance across arbitrary downstream decision functions. Proposition 2 gives one-sided closure under valuation dominance. Proposition 3 establishes an identification boundary when indistinguishable latent states require different decisions.

Case (a) requires invariance over admissible finite changes. Case (b) requires finite completion sets and valuation dominance. Case (c) gives an impossibility result only when latent states share an observation but require conflicting decisions.

For mixed channels, guarantees apply only to components satisfying the relevant assumptions. Assigning an inseparable mixed channel to the residual regime is a conservative convention. It does not follow from Corollary 1.

We made these conditions explicit in Section 4.1 and Appendix A.2.

We also corrected Proposition 2.

The submitted version used ordinary completion-set inclusion. This is insufficient when a manipulation also changes the valuation input.

The revised proposition uses valuation dominance. Every clean completion must have a manipulated completion that is no better for the buyer.

For nonempty finite completion sets, dominance is equivalent to the conservative value not increasing. It prevents rejection-to-acceptance flips for every threshold. If dominance fails, some threshold produces a flip, although the deployed threshold need not.

Ordinary set inclusion is a sufficient special case when shared completions retain their valuation.

We checked dominance by enumerating clean completions and finding corresponding manipulated completions under the specified buyer valuation. This check uses the rendered signals and the buyer's preference parameters. It does not use the hidden test oracle.

Both contract manipulation families satisfy dominance on all 2,605 held-out tasks. Ordinary set inclusion holds on only 486 tasks because the manipulations can also change the headline price.

The revised proof is in Appendix A.1. The experimental verification is reported in Appendix K.

The guarantee is one-sided. It is independent of the learned buyer model only when the deployed decision follows the specified conservative threshold rule. It is not the arbitrary-function invariance of Proposition 1.

2. Comparison with CaMeL.

We ran the official CaMeL implementation in the same AgentDojo benchmark. The compared arms use matched tasks, attacks, model snapshots, and evaluation verdicts (Appendix Q, Table 17).

On closable banking goals, canonicalization and CaMeL both achieve zero measured ASR. Their task utility is 0.414 and 0.422, respectively.

On closable travel goals, both also achieve zero measured ASR. Task utility is 0.475 for canonicalization and 0.025 for CaMeL under gpt-4o-mini. A gpt-4o replication gives 0.600 and 0.250, respectively, with zero measured ASR for both methods.

Slack is different because its security verdict uses attempted calls. Canonicalization blocks the guarded operation but retains a measured ASR of 0.333.

CaMeL also blocks goals outside our selected operator's scope. Its architecture uses additional trusted control information and a restricted execution environment. Some runs fail to complete the legitimate task. Zero measured ASR therefore does not always imply successful task execution.

The comparison shows different coverage and task-utility costs. It does not establish that either method is universally preferable.

We have clarified these assumptions and limits in the revised manuscript.

Thank you for the substantial effort during the rebuttal. The rebuttal has addressed most of my concerns, especially through additional experiments and clarifications. The proposed criterion is valuable, but the demonstrated protection remains limited to well-specified channels under particular information assumptions. After considering the rebuttal together with the other reviewers’ comments, I maintain my original score. The initial submission also has room for improvement in presentation. I believe this work has potential and encourage the authors to further develop it. Thank you again.

We thank the reviewer for the careful reading and for acknowledging that the rebuttal addressed most of the concerns.

We agree with the remaining point, and would add one observation about how we read it. Exactness is always relative to a specified set of admissible actions; that is precisely what makes it checkable in advance rather than after a failure. The alternative is a defense that appears unrestricted, and our added probe measures what that buys: a model that refuses every injection-shaped payload at 0.000 still transfers to the attacker in 0.567 [0.392, 0.726] of episodes once the same account is presented as a correction to the document’s own payee, while the structural operator refuses that transfer under all six framings because it never reads the framing. We therefore regard the specification requirement as a property of the problem that the criterion makes explicit, rather than a restriction peculiar to this method.

On presentation, the revision states the scope condition in the abstract and reorganises the introduction around this distinction, with the supporting competence analysis moved out of the main text.

We thank the reviewer again for the encouragement.

Reviewer_JG1XOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to ensure that buyer agents based on Large Language Models (LLMs) remain invariant to strategic "rephrasings" by sellers while maintaining sensitivity to genuine concessions offered by sellers. The authors propose a set of criteria covering three types of scenarios: (1) filtering out information irrelevant to value; (2) applying a worst-case assumption to enumerable scenarios; and (3) estimating probabilities for unverifiable information that cannot be directly accessed. The authors evaluated these preprocessing operators across multiple negotiation benchmark datasets and the AgentDojo platform with seven different families of LLMs.

Critical strengths

S1. The authors provide a clear taxonomy of information processing. Furthermore, empirical testing demonstrates that the safety-prompting strategies employed by various models differ significantly and, in most cases, fail to provide robust protection, which demonstrating the necessity for deterministic preprocessing.

S2. This paper draws a clear distinction between "framing resistance" and "concession sensitivity." Additionally, by applying "class-two operators" within the AgentDojo framework, the authors successfully reduced the success rate of "on-channel attacks" to zero.

Critical weaknesses

W1. The results are a predetermined inevitability: the method requires the manual, a priori specification of particular operational channels AkA_k and representations ϕ\phi, rendering it incapable of discovering novel, unforeseen natural language attack vectors.

W2. The work lacks technical contribution and is limited in its scope of evaluation: Proposition 1 is essentially a trivial statement, while the "class-two operator" is functionally equivalent to a standard "allowlist" mechanism which is similar to CaMeL-style approaches.

Questions for authors

Q1. How can these operators handle seller manipulations in unconstrained natural language prose where AkA_k cannot fit cleanly into structured fields?

Q2. What is the fundamental difference of the class-two closure operator from CaMeL-style approaches?

2 comments

We added a direct CaMeL comparison and clarified the scope of the attainability criterion.

W2/Q2. How does the method differ from CaMeL?

The AgentDojo implementation uses an allowlist over trusted structured fields. We do not claim that this mechanism is new.

Our contribution is a criterion for a specified channel and representation. It identifies the conditions required for a structural guarantee. Section 5.4 tests its predictions by changing the representation while holding the underlying manipulation fixed.

We ran the official CaMeL implementation in AgentDojo 0.1.35, using suites v1.2.1 and the same gpt-4o-mini-2024-07-18 snapshot across arms. We used AgentDojo's security and task-utility verdicts (Appendix Q, Table 17).

The table reports attack success rate (ASR) and task utility.

Suite and goal classCanonicalizationCaMeL
Banking, closable (128).000 / .414.000 / .422
Travel, closable (40).000 / .475.000 / .025
Banking, class three (16).438 / .562.000 / .312
Travel, class three (100).300 / .390.000 / .050

Both methods achieve zero measured ASR on the tested closable banking and travel goals. Utility is similar in banking. Canonicalization retains higher task utility in travel under this configuration.

We repeated the closable travel subset with gpt-4o-2024-08-06. Both methods again achieve zero measured ASR. Task utility is 0.600 for canonicalization and 0.250 for CaMeL over 40 pairs.

CaMeL also blocks attacks outside the scope of our selected operator. This is consistent with its broader execution architecture. CaMeL uses a trusted user query, a restricted execution language, and explicit security policies. Some runs fail to complete the legitimate task. Zero measured ASR must therefore be considered alongside task utility and execution failures.

Slack provides another qualification. Its closable goal is scored from attempted calls. Canonicalization blocks the guarded operation, but its measured ASR remains 0.333.

The comparison uses matched benchmark conditions. The methods do not have identical trusted information. We do not claim universal superiority over CaMeL.

W1/Q1. What about prespecified channels and natural-language attacks?

The criterion takes an action set, a representation, and a utility-relevance model as inputs. It does not discover arbitrary attacks from unrestricted prose. Actions outside the specified set are outside the guarantee.

The criterion makes predictions that can be tested before evaluating the downstream buyer.

Section 5.4 provides such a test. Two matched manipulations have similar undefended effects, +0.242 and +0.237. A public arithmetic identity exposes the inconsistency in one manipulation. Its measured effect falls to zero under closure. The other manipulation preserves the identity and remains effective at +0.237.

Appendix A.9 evaluates free-text seller messages. Manipulation effects persist for three of four tested model families. Applying closure to unseen prose still requires a correct semantic representation. We have not established reliable parsing of unseen phrasings.

Regarding Proposition 1, its proof is short because it quantifies over arbitrary downstream decision functions. Distinct preprocessed outputs can be ordered unfavorably by some such function. Exact representation invariance is therefore necessary for that model-uniform guarantee.

A fixed known model may admit weaker conditions. Proposition 2 concerns a different setting. It establishes one-sided protection for the specified conservative threshold rule under valuation dominance. We have made this distinction explicit in Section 4.

Thank you for the detailed rebuttal and the extra comparison with CaMeL. The results make sense to me and help me understand the position of the paper accurately. It would be helpful to include these results in the revised paper. Most of my previous concerns have been addressed, so I have increased my score. Best of luck!

Reviewer_DcQtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

An AI buying agent sees a seller's listing and decides accept/walk. The seller doesn't have to change the actual deal to flip the agent's mind — just reword it, bury fine print, add fake urgency. The paper asks: how do you make a rule that doesn't move on re-framing but still moves on real price concessions? They give an ex-ante criterion: look at what a manipulation channel can change vs. what actually carries value, and it sorts into three points:— (1) pure presentation, just drop it for free; (2) payoff uncertainty you can enumerate, take the worst case over completions; (3) latent stuff like unverified quality, provably unclosable, just estimate it. The key theorem, because the buyer is a black-box LLM you can't redesign, a universal guarantee forces exact invariance, not just "manipulation never helps" — which is why prompt defences swing wildly across models. On CraigslistBargain, closure drops harmful acceptance 0.097→0.037 at no utility cost, closed channels move by exactly 0.0000, and real concessions still get accepted. In AgentDojo, on-channel attacker success goes 0.465→0.000.

Critical strengths

The experimental design is solid and correct: hold the deal fixed, vary the rendering, test concessions separately — so "reject everything" can't game the metric. Proposition 1 gives the real insight: black-box deployment collapses the old cone relaxation into exact invariance, and the 17-fold spread in a single safety prompt across models backs it up. They make falsifiable predictions (re-encoding latent quality splits it into a closable half and an open half) and both halves land exactly as predicted. Statistical hygiene is solid — paired bootstraps, Holm correction; they report missed opportunities and regret alongside harm.

Critical weaknesses

The exact 0.0000 only holds on channels they already enumerated. A strategic seller who reads the paper will break the completion-set inclusion condition, and the authors admit best-response inflates residual risk "several-fold." The channel that actually matters most — "is the item really as good as claimed?" — is the unclosable third bucket, handed to a DRO layer whose coverage fails on ~25% of messages. So the clean results are on the easy channels. LLM numbers are small-n (n=200/family), and the headline +0.270 is one model on a subset. AgentDojo "success" is mostly because banking has 8/9 goals on the closable channel; travel and Slack barely move because most goals are class-three.

Questions for authors
  1. Of real-world harmful acceptance, how much is on closable channels vs. the unclosable latent-quality channel?
  2. Once the seller best-responds to the published closure, what's the adaptive cost to keep exactness?
  3. Can a model-family-specific cone relaxation beat exact invariance in practice, or is exactness unavoidable?
1 comment

We added an adaptive seller experiment to address Q2. We also measured channel coverage and clarified the scope of the model-independent guarantee.

Q2. Does closure survive a seller who knows the rule?

We tested an adaptive seller on 2,605 held-out CraigslistBargain tasks. The seller knows the closure rule and chooses among 34 candidate renderings. The true contract and expected settlement remain fixed.

Seller actionDominanceClosure HARReject-to-accept flips
Clause additions (16)Holds0.00000
Clause swaps (15)Fails0.0000155
Concealment, at most 7%Fails0.0472278
Concealment, at most 21%Fails0.1274793

Two results matter. First, no rejection-to-acceptance flips occur under the tested actions satisfying valuation dominance. This agrees with Proposition 2. Second, concealment breaks dominance and produces harmful acceptance. Its HAR matches that of the unprotected buyer in these experiments.

Clause swaps also break dominance, but produce no harmful acceptance in this construction. They change decisions without making the newly accepted offers harmful. We clarified this distinction in Appendix K, Table 16.

Proposition 2 guarantees that a manipulation satisfying dominance cannot change rejection into acceptance. In this benchmark, clean harmful acceptance is zero for the scripted policies. The one-sided guarantee therefore prevents an increase in harmful acceptance under the tested dominant manipulations.

What does additional protection cost?

We tested two approaches under additional assumptions.

An accurate trusted reference contract restores the binding information. In our construction, this adds no decision-conservatism cost. It does require a trusted information source.

Alternatively, a known upper bound on concealed contract mass permits a conservative price adjustment. Under the tested price model, adjustments of 7% and 21% of the listing price provide a no-flip guarantee for the corresponding concealment bounds. These adjustments reduce valid acceptance substantially.

Smaller adjustments of 4% and 10% eliminate harmful acceptance on the evaluated tasks, but provide no such guarantee.

Neither approach establishes protection against arbitrary concealment. We now state this limitation in Appendix K and Appendix N.

Q1. How much harmful acceptance is closable?

We cannot estimate real-world channel prevalence from these benchmarks.

In the new AgentDojo evaluation, 189 of 389 task and injection pairs belong to goals classified as closable. These account for 97 of 183 successful undefended attacks under the fixed model snapshot. The distribution varies across suites. Banking contains 8 closable goals out of 9, travel 2 out of 7, and Slack 1 out of 5 (Appendix Q).

In the economic benchmark, contract complexity and tail incentives increase harmful acceptance by 0.093 and 0.127. Closure reduces both effects to 0.000. State misrepresentation and selective disclosure remain open, with effects of 0.086 and 0.048 (Table 14).

These measurements describe the evaluated benchmarks. They do not establish how frequently these channels arise in deployed markets.

Q3. Could a model-specific relaxation retain more utility?

Yes. Proposition 1 requires exact representation invariance only when protection must hold for every downstream decision function.

A fixed model with known behavior may admit weaker conditions. Such a relaxation can retain more information. We now make this distinction explicit after Corollary 1.

Our guarantee concerns the model-uniform setting. We do not claim that exact invariance offers the best utility tradeoff for every fixed buyer.

Regarding the +0.270 result, it belongs to one model family under the combined attack. We retain that qualification. The new adaptive experiment covers all 2,605 scripted test tasks. The structural guarantee does not depend on the LLM sample size.

Comments on the paper

1 comment

We have completed additional experiments and revised the manuscript to address the main concerns.

Adaptive sellers. We tested a seller that knows the closure rule and selects from 34 candidate renderings on 2,605 held-out tasks. The results confirm one-sided protection under valuation dominance. They also show how concealment breaks the guarantee and measure the cost of restoring protection under a known concealment bound (Appendix K, Table 16).

Comparison with CaMeL. We ran the official CaMeL implementation on matched AgentDojo tasks, attacks, and model snapshots. We report attack success, task utility, and execution failures. We also explain the differences in trusted information between the two methods (Appendix Q, Table 17).

Theoretical clarification. We added Corollary 1 to distinguish the local geometric classification from the conditions required for operational guarantees. We also corrected Proposition 2. The revised condition, valuation dominance, accounts for changes in both the completion set and the valuation input. We verified it directly on the evaluated contract attacks.

During this audit, we found that an earlier +0.014 harmful-acceptance result for incomparable contract sets was not reproduced by the experiment logs. The recorded change is 0.0000. We corrected Appendix K.

The responses below provide the results and clarify the scope of each guarantee.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.